On Thu, 11 Apr 2013, Tejun Heo wrote: > On Thu, Apr 11, 2013 at 12:52:03PM -0700, Tejun Heo wrote: > > If this becomes an actual bottleneck, the right thing to do is making > > css ref per-cpu. Please stop messing around with refcounting. > > If you think this kind of hackery is acceptable, you really need to > re-evaluate your priorities in making engineering decisions. In > tightly coupled code, maybe, but you're trying to introduce utterly > broken error-prone thing as a generic block layer API. I mean, are > you for real? > > -- > tejun All that I can tell you is that adding an empty atomic operation "cmpxchg(&bio->bi_css->refcnt, bio->bi_css->refcnt, bio->bi_css->refcnt);" to bio_clone_context and bio_disassociate_task increases the time to run a benchmark from 23 to 40 seconds. Every single atomic reference in the block layer is measurable. How did I measure it: (1) use dm SRCU patches (http://people.redhat.com/~mpatocka/patches/kernel/dm-lock-optimization/) that replace some atomic accesses in device mapper with SRCU. The patches will likely be included in the kernel to improve performance. (2) use the patch v2 that I posted in this thread (3) add bio_associate_current(bio) to _dm_request (so that each bio is associated with a process even if it is not offloaded to a workqueue) (4) change bio_clone_context to actually increase reference counts: static inline void bio_clone_context(struct bio *bio, struct bio *bio_src) { #ifdef CONFIG_BLK_CGROUP BUG_ON(bio->bi_ioc != NULL); if (bio_src->bi_ioc) { get_io_context_active(bio_src->bi_ioc); bio->bi_ioc = bio_src->bi_ioc; if (bio_src->bi_css && css_tryget(bio_src->bi_css)) { bio->bi_css = bio_src->bi_css; } bio->bi_flags |= 1UL << BIO_DROP_CGROUP_REFCOUNT; } #endif } (5) add "cmpxchg(&bio->bi_css->refcnt, bio->bi_css->refcnt, bio->bi_css->refcnt)" to bio_clone_context and bio_disassociate_task Now, measuring: - create 4GiB ramdisk, fill it with dd so that it is allocated - create 5 nested device mapper linear targets on it - run "time fio --rw=randrw --size=1G --bs=512 --filename=/dev/mapper/linear5 --direct=1 --name=job1 --name=job2 --name=job3 --name=job4 --name=job5 --name=job6 --name=job7 --name=job8 --name=job9 --name=job10 --name=job11 --name=job12" (it was ran on 12-core machine, so there are 12 concurrent jobs) If I measure kernel (4), the benchmark takes 23 seconds. For kernel (5) it takes 40 seconds. Mikulas -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel